Microsoft AI has introduced seven new models built internally, marking a broader push to develop its own foundation models and the systems around them. The company said the release spans image, voice, transcription, coding and reasoning, and is part of a larger effort it says will shape the next phase of AI.
In announcing the models, Microsoft AI chief Mustafa Suleyman framed the effort as a response to rapid growth in compute and model capability. He said the company expects a major increase in training scale over the next few years and described Microsoft’s goal as building systems that can keep improving as the frontier moves forward.
The new family includes MAI-Thinking-1, a reasoning model Microsoft says performs strongly in its size class on software engineering and mathematics tasks. It also includes MAI-Code-1-Flash, a coding model aimed at agentic use inside GitHub Copilot, Visual Studio Code and other Microsoft products. Microsoft said the coding model is designed to be efficient and lower cost.
The other models in the lineup focus on media and language tasks. MAI-Image-2.5 handles text-to-image generation and image editing. MAI Transcribe-1.5 is positioned as a transcription model with support for terminology across 43 languages. MAI-Voice-2 generates speech in 15 languages and can adapt to a speaker from a short sample. Microsoft said a lower-cost Flash version of the voice model is coming soon.
The company also said the models will be available through its own Foundry platform and, in some cases, through outside services such as OpenRouter, Fireworks and Baseten. Microsoft said developers will be able to fine-tune the weights of the models themselves, which it presented as a first for its offerings.
Beyond the models themselves, Microsoft used the announcement to introduce Frontier Tuning, a workflow-based approach that uses reinforcement learning in real-world environments. The company said the method is designed to adapt AI to the specifics of a given organization’s processes, using the actual traces of work completed by agents.
Microsoft described these reinforcement learning environments as private training spaces that allow a model to learn from an organization’s own data and actions. The company said the approach keeps that knowledge within the customer’s control and can improve both performance and efficiency.
Microsoft pointed to internal and customer examples to support the claim. It said a tuned model for Excel matches GPT 5.4 while being up to 10 times more efficient. It also said an enterprise customer using a Frontier Tuning setup achieved the highest win rate among models tested at roughly 10 times lower cost.
The pitch is aimed at businesses that want AI systems tailored to specific tasks rather than general-purpose behavior. Microsoft said the approach is intended to let organizations build models on their own data, within their own environment, while keeping oversight in place.
Microsoft also announced a collaboration with Mayo Clinic on a frontier AI model for healthcare. The model will combine Mayo Clinic’s clinical expertise and de-identified data with Microsoft’s AI capabilities.
According to Microsoft, the healthcare model will be built to handle broad clinical reasoning and healthcare use cases that current general-purpose systems cannot match. The first deployment will take place inside Mayo Clinic’s environment, with the company expecting it to support diagnosis and treatment planning. If validated, Microsoft said the model could later be made available to other organizations through Foundry.
Microsoft said the model will be owned by Mayo Clinic, a structure it said is meant to reinforce trust, safety and responsible handling of clinical data.
The announcements reflect Microsoft’s broader effort to build a vertically integrated AI stack, from models and infrastructure to product deployment and custom tuning. The company said its long-term aim is to create systems that remain under human control while becoming more capable and more tailored to the people and organizations that use them.